The process package's doc names the proof, so the SDK publishes a new version that node-tools, its importer, is built against in the same merge call.
Builds: design 47 H §5 step 7, G §370
The views test read NumAckPending as soon as the picture and the adopted handler had said the last event, but the
reader acknowledges after both, and the bus takes acks on its own; on the loaded build node (pipeline main.54) one
ack was still pending. The reader binds the module's durable once (the hub's one pull); the guard of H-A-196 is not
involved. The test now publishes a mark the hub hands after every event before it, acknowledged with the bus's
answer, and reads that the ack floor reached the mark; an event the owner says after it may still wait.
Builds: design 47 H §774, H-A-196, H-A-197
The tool runner would otherwise copy the claim's wrapper (review round 1, P7):
ConsumeOnce claims the durable, pulls it and answers the loop, and HearEvents
uses it; Conn.Consume claims through the same guard. A drain keeps its claim
until the loop ended, and the loop's Closed is closed only once the claim is
released, so a bind during a drain is refused and one after it is never.
Builds: design 47 H H-A-196, H-A-197
A second pull loop on a durable the same process already pulls takes every other
event from the first, and nothing said so: the tool runner's two loops on the
indexer bundle's durable left the catalog seat's revision verb stale twice today.
The SDK now refuses it at bind time, naming the durable and both binders:
PullOnce claims a durable (by stream, name and the time the bus made it), and
HearEvents and Conn.Consume claim through it, releasing on stop, drain or the
bus ending the loop.
Builds: design 47 H H-A-196
The hub's hook, the path the live fault took, had no test that failed
without it, and each Start appended it again, so a second Start counted
every event twice. Start now wires it once per hub and keeps the time of
the first start; the shared-durable integration test holds the reader's
finding of the one event nobody takes, counted once across two starts.
Builds: design 47 H H-A-169
An event a module consumes that no handler takes was said in the journal
alone, so a part that never acted on it (the runner's placing at 17:34) was
seen only by reading the log. The event hub now tells a hook of each such
subject; the process reader, where it holds a findings writer, counts them
in one finding event-unhandled per module (the count since the start and
the last subject) and clears it at its start, since a new build may handle
the event. Without a findings writer the journal line stands alone.
Builds: design 47 H H-A-166, H-A-169
A seat's contract says an event on its bare subject (the seat mesh-catalog's
registration-changed), while both view readers adopted it on `<subject>.>`
alone, so the event hub acknowledged it unhandled and the part never acted.
The process reader now registers one handler on both forms of the hub (the
two patterns do not overlap, so a message reaches it once), and the stdio
reader subscribes both forms with the one handler.
Builds: design 47 H H-A-166
go/callers.go goes (stdio.Callers, VerbCallers, ParseCallers, ParseVerbCallers, ConfiguredVerbCallers, HoldVerbCallers, AnyPerson, CallerSeatPrefix; ModuleOf stays in caller.go). seatlist.NewPart(prefix), Tools(r, read, again); ToolsWith, Refusal, RefusalWith and Part.Holders go. views.FirstViewTool loses its callers argument; views.ParseCallers and views.Refusal go. approval.ViewCallersPrefix and views.CallersEnv stay until the question of how the gate tells a first-view tool is answered.
Builds: design 47 H §752, §764, H-A-41, H-A-66
A definition naming callers is refused (a field the format does not know); Callers, DefaultCallers, EffectiveCallers, CallerEntryProblem and the Caller constants go. The test seat list's seats lose their callers. The seat mesh-delivery-checker's bootstrap, test and captured copies go.
The copies of git, mesh-delivery and mesh-catalog (bootstrap, test, captured) have only their callers lines taken out by hand so this commit's tests are green: they are replaced by mesh-delivery's contracts files in sdk-seat-copies-h1's remainder.
Builds: design 47 H §752, §764 (The checker, Every seat contract), H-A-41, H-A-42, H-A-57
The bootstrapped, test and beside copies are the seats' files as the contracts hold them after the removal of the pull
request path (design 47 G); the views test names a
state the indexer still has.
Builds: ADR 0681 §7 (mesh-sdk's seat list definitions of mesh-config's `kept` and the mesh-catalog seat's `config-candidates` go).
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Cause: startRead looked only in the definitions the runner holds. The seat list is itself read through the seat
mesh-catalog, so before the catalog half is ready the runner holds none, and every first view it waits on (its own
and another runner's, from any account) was refused with "marks not read yet" for ever.
Quick fix: let the one account that stalled through (calledByModule). Thorough fix (chosen): the foundation's bootstrapped
definitions answer for the seats they name wherever the held list lacks them, so a first view runs whoever asks; an act
of the same seat and a seat no definition names stay held.
Test: TestTheFirstViewsRunBeforeAnySeatListIsRead (fails without run.go change).
Builds: design 47 G A-387, A-388, A-433
A-435: the mesh MCP server's account, a person and an unproven caller are held as before while the definition is unread;
a module's account runs it (A-433).
Builds: design 47 G A-433, A-435
The assigner's first view of the indexer's modules stayed behind from 2026-11-14 15:04 UTC: events said the state was
made 2026-10-29, the answer named no make time, and zero is always before it, so it asked 378 times and took in no
event; every want kept its build of that day (the indexer's own new build was never wanted, so mesh-delivery #219
was never taken in).
The writer's lock was held through the event's resend (up to PublishRetryFor, one minute), so one unacknowledged event
blocked every other Write and Clear behind it.
Builds: design 47 G §383 (the next statement covers an event not acknowledged; the state alone holds the finding)
The take-in's call of the forge bridge's tool was refused while the runner's catalog half was
unread: a module tool names no class, and its marks live only in the half the take-in makes.
A module tool whose proven caller is another module's account now runs unmarked then, logged
once per tool with its caller; the console's account (the runner's own), a person, the
controller and an unproven caller still wait and are judged as before.
Builds: design 47 G lead answers A-389
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
Before its catalog half is read, a tool runner's gate admitted a hand-kept
list of seat verbs; the first take-in then failed on git.pull and the
runner's own mesh-catalog.seat-definitions. The gate now reads the seat
definition the runner holds (its catalog's Definitions, the seat list):
a verb classed read explicitly and not marked needs-approval runs; a
verb naming no class, an act, or a marked read is judged as before.
Builds: design 47 G lead answers A-387, A-388
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The first take-in reaches the seat git before a tool runner's seat list is
read, so its definition is bootstrapped, a copy of its contracts file with
its row in the bootstrap test; the catalog and broker copies follow their
contracts, which class their reads.
Builds: design 47 G lead answers A-387, A-388
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The tool is told by the callers list its bundle's environment carries (views.CallersEnv); the list still limits who.
Builds: design 47 G A-377 (widens A-359 toward A-353)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h